Framsticks

Thanks, I will try to do it in this way.

I still have one question left.

I'm trying to make a creature that behaves in a biological way by perceiving
its own internal state like being fertile or hungry.
Now I'm not sure about how to model this. I could make a sensor connected to
the neural network that simply sets its state to the fertility variable or I
could add the vertility variable to the getInputState of each neuron since
this is an internal variable that needs no perceiving.

This is in essence of course the same approach, but I was wondering which
approach is computationally more attractive. A third, but not biological
solution is to hard-code the rules like if(fertile) than search mate.

Regards,

Tom Schut

"Maciej Komosinski"
schreef in bericht news:e84ddf$sd5$1@cancer.cs.put.poznan.pl...
> Hi Tom,
>
>> I am working on a project in framsticks and want to develop learning
>> creatures. The problem with this is that in any type of learning,
>> especially hebbian, there has to be an interaction between neurons to
>> change the weights. So my question is: is it possible to get a reference
>> to an input neuron AND change the weights from that neuron to your own. I
>> tried looking a the NeuroProperties object, but it has no documentation.
>
> Let me quote my answer to Matej from Prague who was also interested
> in this issue:
>
>
> > And I have one technical question. I would like to use different types
> > of
> > synapses, to allow learning (change of synapses, i.e. weights) in an
> > individual (different types of Hebbian learning) and to evolve just
> > parameters of the synapses (e.g. excitatory/inhibitory, method of
> > learning...). Are such things theoretically possible (if I use the
> > script), or the software allows only simple weights between neurons,
> > that
> > do not change in an individual?
>
>
> Weights themselves cannot change (as encoded in the genotype), but
> you can do virtually anything in a neuron that you define in the *.neuro
> script file, i.e. you can define which neuron parametes whould be evolved
> and which should not, and how these parameters affect neuron
> characteristics. So you can easily overcome the constraint of constant
> weights - for example, all weights could be always "1", and your
> "dynamic" weights would be a property of a neuron.
>
>
> > 1)How to make the neurons working? I can make my own neuron definition,
> > where I can edit properties. The synapses can be at the target
> > (postsynaptic) neuron. However, I need a synapse for every input neuron.
> > So I need something like a dynamic array and I also need access to the
> > input neurons (I have to know if it did not fire as well, if rule
> > presynaptic). Or can I somehow control everything from the onStep) of
> > the
> > experiment definition?
>
> You can have as many local variables/objects as you need in each
> neuron. For example, you can create an array with the size equal
> to the number of inputs in each neuron. *See the scripts/delay.neuro
> file*.
>
> You can get (unweighted) inputs using the getInputState function.
>
> > 2) I hope I studied all available documentation, but I am missing some
> > description of relations between the classes, what is loaded when and
> > where from... Is there perhaps a documentation for programmers? I have
> > read the Scripting presentations, but that does not really uncovers the
> > data flow between the objects... For instance I am not sure where Neuro
> > and NeuroClass live and how can I get to them... I would also like to
> > know, how the activations spread in the nn, is it synchronous or
> > asynchronous...
>
> It is synchronous. Each neuron sets its own output state.
>
> As to the data flow... in Framsticks, you make event handlers
> or dynamically create objects, or use "static" objects. Some
> objects only exist (are accessible) in some specific contexts,
> shown as filled rectangles in the left frame at
> http://www.framsticks.com/common/script/docs/index.html
>
> > 3)Do my properties of synpases get automatically evolved under f1, if
> > added to neuron definition (property:)? How will they be encoded,
> crossed-over
> > etc.? Or do I have to create my own encoding?
>
> Yes; simple, published (i.e. not those with flags:32) properties are
> automatically mutated (simple are integer or floating point properties).
> String properties or properties of other types are not mutated.
>
>
> MacKo